Does Education Pay in Urban China- Estimating Returns to Education Using Twins

7/28/2019 Does Education Pay in Urban China- Estimating Returns to Education Using Twins

http://slidepdf.com/reader/full/does-education-pay-in-urban-china-estimating-returns-to-education-using-twins 1/39

Does Education Pay in Urban China?

Estimating Returns to Education Using Twins∗

Hongbin Li Pak Wai Liu Ning Ma Junsen Zhang†

September 23, 2005

∗We would like to thank Mark Rosenzweig for his advice in the survey work, Larry Katz for his comments,and the Hong Kong Research Grants Council for funding the project. Junsen Zhang also thanks NIHHD046144 for partial financial support. The usual disclaimer applies.

†All of the authors are affiliated with the Department of Economics of the Chinese University of Hong

Kong, Shatin, Hong Kong. Corresponding author: Junsen Zhang, Tel.: 852-2609-8186; fax: 852-2603-5805;E-mail: [email protected]



Does Education Pay in Urban China?Estimating Returns to Education Using Twins

Abstract

This paper empirically estimates the returns to education using twins data that theauthors collected from urban China. Our ordinary least-squares estimate shows thatone year of schooling increases an individual’s earnings by 8.4 percent. However, oncewe use the within-twin-pair fixed effects model, the return is reduced to 2.7 percent,which suggests that much of the estimated returns to education in China that have beenfound in previous studies are due to omitted ability or the family effect. This findingsuggests that well-educated people are faring well in China mainly because of theirsuperior ability or family background advantages, rather than because of knowledgethat they acquired at school. We further investigate why the true return is low andthe omitted ability bias high, and find evidence that it may be a consequence of thedistinct education system in China, which is highly selective and exam oriented. Morespecifically, we find that high school education mainly serves as a mechanism to selectcollege students, and has zero returns in terms of earnings. In contrast, both vocationalschool education and college education have a large return that is comparable to thatfound in rich Western countries.

JEL Classification : J31; O15; P20



1 Introduction

Although estimating the return to education has been an important econometric exercise

since the seminar work of Mincer (1974), only recently have economists begun to estimate

it using Chinese data. Several studies that draw on data from urban China from the 1980s

and 1990s find rather low returns, with one year of schooling increasing earnings by only 2-4

percent (Byron and Manaloto, 1990; Meng and Kidd, 1997). This finding has caught the

attention of many labor economists, including James Heckman, who generally think that the

estimates of the return to education in China were formerly low because most of the urban

economy was still under a planned regime in the 1980s and 1990s. However, they believe that

the return should have increased after more than two decades of economic transition from

a planned regime to a market regime, as in market economies a large gradation in earnings

according to the level of education reflects the return to the investment of individuals in

education (Mincer 1974; Becker 1993).1 Recent data have shown that the return to education

has indeed risen in China. Heckman and Li (2004) find that the return to education had

risen to 7 percent by 2000. Using a repeated cross-sectional dataset of a 14-year period

(1988-2001), which is the best large-scale dataset of this kind, Zhang et al. (forthcoming)

find a dramatic increase in the return to education in urban China from only 4 percent in

1988 to more than 10 percent in 2001.

Despite the rapid accumulation of evidence on the return to education in China, no

study has yet established causality. An ordinary least-squares (OLS) estimation of the effect

of education on earnings cannot prove causality, because well-educated people may have

high earnings as a result of their greater ability or better family background. In other words,

education may be correlated with unobserved ability or family background, which would

1In fact, this assertion has contributed to a lively debate among sociologists who study institutionaltransformation and social stratification in former state socialist societies (Rona-Tas, 1994; Bian and Logan,1996; Parish and Michelson, 1996; Szelenyi and Kostello, 1996; Walder, 1996; Xie and Hannum, 1996; Gerber

and Hout, 1998; Zhou, 2000).



make any correlation between education and earnings spurious. Because of the difficulty

in breaking endogeneity due to unobserved ability, the true return to education in China

remains elusive.

Our first goal in this paper is to empirically measure the causal effect of education on

earnings by using twins data that two of the authors collected in urban China. As is argued in

the literature (Ashenfelter and Krueger, 1994; Miller et al., 1995; Behrman and Rosenzweig,

1999; Bound and Solon, 1999; Isacsson, 1999),2 as monozygotic (from the same egg) twins

are genetically identical and have a similar family background, the effects of unobserved

ability or family background should be similar for both twins. Thus, taking the within-twin-

pair difference will, to a great extent, reduce the unobservable ability or family background

effects that cause bias in the OLS estimation of the return to education. Intuitively, by

contrasting the earnings of identical twins with different years of education, we can be more

confident that the correlation that we observe between education and earnings is not due to

a correlation between education and an individual’s ability or family background.

Our empirical work shows that most of the effect of education on earnings from the OLS

estimates is actually due to the effects of unobserved ability or family background. Our OLS

estimate shows that the return to one more year of education is 8.4 percent, which is close

to other recent estimates that use Chinese or Asian data (see, for example, Psacharopoulos,

1992; Heckman and Li, 2004; Zhang et al., forthcoming). However, once we use the within-

twin-pair fixed effects model, the return is reduced to 2.7 percent, which suggests that much

of the estimated return using the OLS model is due to the omitted ability or the family

effect. In other words, education in China is more important for selecting people of high

ability to progress through the system than it is for providing knowledge or training that

2The earliest attempt to look at siblings data in economics can be traced back to the dissertation of Gorseline (1932). Not satisfied with siblings data, economists started to use twins data in the late 1970s,when the work of Behrman and Taubman (1976), Taubman (1976a, 1976b), and Behrman et al. (1977) waspublished. The interest in using twins data was recently revived with the work of Ashenfelter and Krueger

(1994) and Behrman et al. (1994).



will enhance earnings. This finding is confirmed by a generalized least-squares estimation

that includes the co-twin’s education as a covariate.

Thanks to the new advances in twins studies that have been made by Ashenfelter and

Krueger (1994), Ashenfelter and Rouse (1998), Bounjour et al. (2003), Hertz (2003), and

others, we have been able to obtain good-quality data and address several problems that

are inherent in twins studies. First, our correlation tests that follow Ashenfelter and Rouse

(1998) show that the between-family correlations of education with other family character-

istics are all larger in magnitude than the within-twin-pair correlations, which suggests that

the within-twin-pair estimate of the return to education may be less affected by omitted

variables than the OLS estimate. Second, we address the potential bias that is caused by

the measurement error in the education variable by using the instrumental variable approach

of Ashenfelter and Krueger (1994). After the correction of measurement error, the estimated

return to education rises by about one percentage point to 3.8 percent.

The low estimated return to education and high selectivity (or ability bias) differ

sharply from evidence from twins data from other countries (see, for example, Ashenfelter

and Rouse (1998), Behrman and Rosenzweig (1999) and Bonjour et al. (2003)). Our second

goal in this paper is thus to ascertain what is so different about China. Although the

remaining features of a planned economy could be used to explain the low return, we provide

an alternative explanation in this paper. We argue that the low return and high selectivity

may be a consequence of the distinct education system in China. Because of the huge

population awaiting an education and the limited number of college (and university) places,

entrance to college is extremely competitive. The Chinese solution to this is examinations.

Only the very talented can score high enough in the college entrance exams to advance to

higher education, and thus non-tertiary education, and in particular high school education

and the associated entrance exams, has become a very important selection mechanism. This



explains why the ability bias is so high in our OLS estimates. Moreover, to prepare students

for college entrance exams, non-tertiary education in China, and in particular high school

education, is totally exam oriented, and thus adds little value in terms of general knowledge

or workplace skills. As a result, such exam-oriented high school education has a low return,

which has also dragged down the overall return to education.

The twins data that we have collected allow us to test whether the Chinese education

system should indeed be blamed for the low return to and high selectivity of education.

To this end, we estimate the returns to different levels of education by using OLS, within-

twin-pair, and IV estimations. Arguably, exam-oriented high school education should have

the lowest return among all of the education levels, and the final-stage education levels,

such as vocational and college education, should have higher returns because they are less

exam oriented. Interestingly, these hypotheses are confirmed. We find that the return to

high school education is almost zero, but that the return to college education is very large.

According to our estimates, which control for omitted variable and measurement error biases,

college graduates earn 40 percent more than those who have not been to college or vocational

school. These findings suggest that going to high school does not pay unless an individual is

also able to obtain a college degree. Moreover, although the return to high school education

is zero, there is a large return to vocational school education. The return to vocational

school education is as large as 22 percent, or 7.3 percent per year of schooling.

The idea of using twins data to control omitted ability bias excited many labor economists

when it first came out, but its popularity waned when many twins studies found that the

OLS estimates did not differ much from within-twin-pair estimates that controlled for omit-

ted ability. Part of the reason for the low omitted ability bias in previous studies is that

most of these studies draw on data from rich Western countries, where education is not very

selective. To the best of our knowledge, this is the first study of the return to education that



draws on twins data from China, and is probably also the first to draw on Asian twins data.

The education systems of Asian countries, and especially East Asian countries and regions

such as Japan, South Korea, Taiwan, Hong Kong, and mainland China, share similar fea-

tures in that they all have very serious college entrance exams. Understandably, high schools

in these countries or regions place a lot of emphasis on exam-taking techniques, and thus

education may be more selective in these countries or regions than it is in Western countries.

In this sense, twins studies, which largely separate the selection effect from the true return

to education, may be more important for these countries than for Western countries. Our

study is also one of the first to use twins data from developing countries. Twins studies in

developing countries are particularly interesting, because the omitted variable bias may be

larger in these countries, where liquidity constraints and family background are likely to be

important determinants of both education and earnings (Lam and Schoeni, 1993; Herrnstein

and Murray, 1994).

Knowing the true return to education is very important for China, which is experiencing

a transition from a planned economy to a market economy. During the transition process,

the Chinese government must reform all of the economic sectors, such as industry, banks,

the medical system, and education. Given the limited resources that are available, the

government needs to set priorities for government expenditure. Our findings suggest that

the true return to one year of schooling is at most 3.8 percent, which may be far below the

return to investment in physical capital. However, the return is not uniform for different

education levels. We find that the return to high school education is zero, and that in terms

of each year of schooling, the return to both a vocational degree and a college degree is high.

Thus, cutting one year from the three years of high school and using the saved resources for

other education levels may increase the overall efficiency of the economy.

The structure of the rest of this paper is as follows. Section 2 describes the estimation



methods that draw on twins data. Section 3 describes the data and the variables. Section 4

empirically measures the return to education. Section 5 explains why the return to education

is low and selectivity high in China. Section 6 concludes.

2 Method

Our empirical work focuses on the estimation of the log earnings equation, which is given as

yi = X iα + Z iβ + µi + i, (1)

where the subscript i refers to individual i, yi is the logarithm of earnings, X i is the set

of observed family variables, and Z i is a set of observed individual variables that affect

earnings, which includes education, age, age squared, gender, marital status, and job tenure.

µi represents a set of unobservable variables that also affect earnings, that is, the effect of

ability or family background. i is the disturbance term, which is assumed to be independent

of Z i and µi.

The OLS estimate of the effect of education in equation (1), β , is generally biased.

This bias arises because we normally do not have perfect measures of µi, which is very

likely to be correlated with Z i. Intuitively, the cross-sectional comparison of the earnings

of workers with different levels of education will not identify the education effect even if

these workers are identical with respect to other observed variables. This is because workers

with different levels of education may differ in other unobserved characteristics that affect

earnings. As discussed in the introduction, well-educated people may be more capable,

motivated, or blessed with an advantageous family background, and if these advantages

are not completely accounted for, then the OLS estimation will pick up the effect of these

variables. It is therefore difficult to ascertain how much of the empirical association between

earnings and education is due to the causal effect of education, and how much is due to

unobserved factors that influence both earnings and education. The omitted variable bias



depends on cov(Z i,µi)var(Z i)

, which summarizes the relationship in the sample between the excluded

µi and the included Z i, which includes education.

Several approaches may be used to tackle this problem of omitted variable bias. The

first approach is to seek richer datasets that can be used to control more extensively for

measures of ability, family background, and the like. The main problem with this approach

is that the controls inevitably remain incomplete. Nonetheless, we take advantage of our

rich dataset and include many control variables to reduce the omitted variable bias.

A second approach to the omitted variable problem is to apply the fixed effects es-

timator to our twins sample. As monozygotic (from the same egg) twins are genetically

identical and have a similar family background, they should have the same µi. Thus, taking

the within-twin-pair difference will eliminate the unobservable ability and family effect µi,

which causes the omitted variable bias in the OLS estimation. Intuitively, by contrasting

the earnings of identical twins with different levels of education, we can ensure that the cor-

relation that we observe between education and earnings is not due to a correlation between

education and a worker’s ability or family background.

The fixed effects model can be specified as follows. The earnings equations for a pair

of twins are given as

y1i = X iα + Z 1iβ + µi + 1i (2)

y2i = X iα + Z 2iβ + µi + 2i, (3)

where y ji ( j = 1, 2) is the logarithm of the earnings of both twins in the pair and X i is the

set of observed variables that vary by family but not between the twins, that is, the family

background variables. Z ji ( j = 1, 2) is a set of variables that vary between the twins.

A within-twin-pair or fixed effects estimator of β for identical twins, β fe is based on

the first-difference of equations (2) and (3):

y1i − y2i = (Z 1i − Z 2i)β + 1i − 2i. (4)



The first difference removes both the observable and unobservable family effects, that is, X i

and µi. As µi has been removed, we can apply the OLS method to Equation (4) without

worrying about bias that is caused by the omitted ability and family background variables.

A third approach to solving the omitted variable bias is to directly estimate both the

bias and the education effect using the approach that was developed by Ashenfelter and

Krueger (1994). This approach also draws on monozygotic twins data. In this approach, the

correlation between the unobserved family effect and the observables is given as

µi = Z 1iγ + Z 2iγ + X iδ + ωi, (5)

where we assume that the correlations between the family effect µi and the characteristics

of each twin Z ji ( j = 1, 2) are the same, and that ωi is uncorrelated with Z ji ( j = 1, 2) and

X i. The vector of the coefficients γ measures the selection effect that relates to the family

effect and individual characteristics, including education.

The reduced form for equations (2), (3), and (5) is obtained by substituting (5) into

(2) and (3) and collecting the terms as follows.

y1i = X i(α + δ ) + Z 1iβ 2 + (Z 1i + Z 2i)γ + 1i (6)

y2i = X i(α + δ ) + Z 2iβ 2 + (Z 1i + Z 2i)γ + 2i, (7)

where ji = ωi+ ji, ( j = 1, 2). Equations (6) and (7) are estimated using the generalized least

squares (GLS) method, which is the best estimator that allows cross-equation restrictions

on the coefficients. Although both the fixed effects and GLS models control for ability, and

can produce unbiased estimates of the education effect β , the GLS model also allows the

estimation of the selection effect γ .

3 Data

The data that we use are derived from the Chinese Twins Survey, which was carried out by

the Urban Survey Unit (USU) of the National Bureau of Statistics (NBS) in June and July



2002 in five cities of China. The survey was funded by the Research Grants Council of Hong

Kong. Based on existing twins questionnaires in the United States and elsewhere, the sur-

vey covered a wide range of socioeconomic information. The questionnaire was designed by

two authors of this paper in close consultation with Mark Rosenzweig and Chinese experts

from the NBS. Adult twins aged between 18 and 65 were identified by the local Statistical

Bureaus through various channels, including colleagues, friends, relatives, newspaper adver-

tising, neighborhood notices, neighborhood management committees, and household records

from the local public security bureau. Overall, these channels permitted a roughly equal

probability of contacting all of the twins in these cities, and thus the twins sample that was

obtained is approximately representative. (The within-twin-pair estimation method that is

used for this study controls for the first-order effects of any unobserved characteristics that

may have led to the selection of twins pairs into the sample). Questionnaires were com-

pleted through household face-to-face personal interviews. The survey was conducted with

considerable care, and several site checks were made by Junsen Zhang and experts from the

National Bureau of Statistics. Following appropriate discussion with Mark Rosenzweig and

other experts, the data input process was closely supervised and monitored by Junsen Zhang

himself in July and August 2002.

This is the first socioeconomic twins dataset in China, and perhaps the first in Asia.

The dataset includes rich information on the socioeconomic situation of respondents in the

five cities of Chengdu, Chongqing, Harbin, Hefei, and Wuhan. Altogether there are 4,683

observations, of which 3,012 observations are from twins households. For the sample of

twins, we can distinguish whether they are identical (monozygotic) or non-identical twins.

We consider a pair of twins to be identical if both twins respond that they have identical

hair color, looks, gender, and age. Completed questionnaires were collected from 3,002

individuals, of which 2,996 were twin individuals and 6 were triplet individuals. From these



3,002 individuals, we have 914 complete pairs of identical twins (1,828 individuals). We have

complete information on earnings, education, and other variables for both twins in the pair

for 488 of these pairs (976 individuals). The summary statistics of identical twins and all

twins together are reported in the first two columns of Table 1.

For the purposes of comparison, non-twin households in the five cities were taken from

regular households on which the USU conducts regular monthly surveys of its own. The

Urban Survey Unit started regular monthly surveys in the 1980s. Their initial samples

were random and representative, and they have made every effort to maintain these good

sampling characteristics. However, their samples have become less representative over time.

In particular, because of an increasingly high (low) refusal rate among young (old) people, the

samples have gradually become biased toward the oversampling of older people over time.

The survey of non-twin households was conducted at the same time as the twin survey,

and the same questionnaire was used. The summary statistics of our non-twins sample are

reported in the third column of Table 1.

Although our within-twin-pair estimation controls for possible sample selection, it is

interesting to compare the identical twins sample to the other samples that we have. To

facilitate such comparisons, we also provide the basic statistics for a large-scale survey that

was conducted by the USU of NBS as a benchmark (henceforth the NBS sample, reported in

column 4 of Table 1).3 Column 1 shows that sixty percent of our identical twins were male,

and on average the twins were 35 years old, had 12 years of schooling, and had spouses who

also had an average of 12 years of schooling. They had worked for an average of 15 years,

and had monthly average earnings of 888 yuan, where earnings include wages, bonuses, and

subsidies. The individuals in the identical twins sample were younger than those in the NBS

sample and also earned less. Finally, individuals in the non-twins sample (column 3) were

3The NBS has been conducting an annual survey of urban households from 226 cities (counties) in Chinasince 1986. It is the best large-scale survey of this kind.



older than those in the NBS sample and the twins samples.

To ensure the good performance of the within-twin-pair estimation of the return to

education, the within-twin-pair variation of education needs to be large enough. We check

the within-twin-pair variation in education and find it to be rather large. Fifty-three percent

of the twin pairs had the same education, 13 percent had one year’s difference in education,

about 10 percent had two years’ difference, and the remaining 24 percent had a difference of

more than two years. These numbers suggest that we have a large variation of within-twin

difference in education, which is good for the fit of the regressions.

4 Returns to Education

In this section, we report the estimated return to education using different samples and

methods. We start with the OLS regressions using the whole sample, including twins and

non-twins, and then conduct the same OLS estimation using the monozygotic twins sample

and compare the estimated coefficients to those that are estimated using the whole sample.

This comparison may serve as a way to check the representativeness of the monozygotic

twins sample. We then conduct the within-twin-pair fixed effects and GLS estimations using

the twins sample, followed by examinations of possible bias in fixed effects estimates and the

impact of measurement error.

4.1 OLS Regressions Using the Whole Sample

In the first two columns of Table 2, we report the results of the OLS regressions using the

whole sample, including both twins and non-twins. The dependent variable is the logarithm

of monthly earnings. The t-statistics are calculated using robust standard errors. In column

1, we show a simple regression with education, age, age squared, gender, and city dummies

as independent variables. This simple regression shows that the return to education is quite

large. One more year of schooling increases an individual’s earnings by 6.7 percent, which



is quite precisely estimated with a t-statistic of 16.71. The positive coefficient of age and

the negative coefficient of age squared are both significant at the 10-percent level. Earnings

increase with age before 55 years of age, but start to drop after that. The gender difference

in earnings is quite large in urban China, with men having 21.7 percent higher earnings than

women.

When we add other control variables in the second column, including marital status

and tenure, the estimated coefficient of education remains unchanged, which suggests that

omitting these variables results in no bias in the estimated return to education. We do

not find a marriage premium in the sample, as the marriage dummy is not significant at the

conventional level. Job tenure has a positive effect, with one more year at the post increasing

earnings by 1.6 percent.

4.2 OLS Regressions Using the Monozygotic Twins Sample

In this subsection, we repeat the same OLS regressions using the monozygotic twins sample.

Comparing the OLS results of the monozygotic (MZ) twins sample with those of the whole

sample is a way to check the robustness of the estimated coefficients using different samples.

As we only use the MZ twins sample, the sample size is reduced to 976 observations (or 488

pairs of twins).

The regression results that are reported in the third and fourth columns of Table 2

suggest that the return to education is larger for our MZ twins sample. The return to

education is 8.2 percent for the simple regression in column 3, and becomes even larger

when other control variables are included in column 4.4 Thus, the OLS estimate of the

return for the twins sample is about 1.5-1.7 percentage points more than that for the whole

sample. The estimated coefficients of most of the other variables are very similar for the two

samples.

4

These OLS estimates are very close to those using the large NBS sample (Zhang et al., forthcoming).



To summarize, the OLS estimate of the return to education is rather large, even after

we control for many covariates. The remaining effect is 0.084 (column 4). However, we still

do not know how much of this effect is the true return to education, and how much is due to

the effects of unobserved ability or family background. We resort to the within-twin-pair and

GLS estimations to remove the unobservables and estimate the true return in the following.

4.3 Within-Twin-Pair and GLS Estimations

In columns 5 and 6 of Table 2, we report the results of the within-twin-pair fixed effects

estimation, or the estimation using Equation (4). As MZ twins are of the same age and

gender, these variables are dropped when taking the first difference.

The within-twin-pair estimation shows that much of the return to education that is

found by the OLS estimation is the result of the effects of unobserved ability or family

background. Note that the within-twin-pair estimate of the return to education is much

smaller than the OLS estimate. Taking column 6 as an example, it can be seen that the

education effect is 0.027, which is only about one third of the OLS estimate using the same

twins sample. This suggests that two thirds of the OLS estimate of the return is actually

the unobserved ability or family effect. Other control variables are not significant in the

within-twin-pair estimation.

We next turn to the GLS estimator for Equations (6) and (7), which can directly

estimate both the return to education and the ability or family background effect. In the

last two columns of Table 2, we report the GLS estimates, including the covariates that are

used in the OLS estimates. In addition to an individual’s own education, we also include

the sum of the education of both twins as an independent variable. The coefficient of this

new variable will be the estimated ability or family effect, that is, γ in Equations (6) and

(7). The GLS model is estimated by stacking Equations (6) and (7) and fitting them using

the SURE model.



The GLS estimation again shows that the return to education is small, whereas the

omitted ability or family effect is large. The coefficients of an individual’s own education are

only 0.025-0.027, which are exactly the same as the values for the within-twin-pair estimates.

The estimated family effect, that is, the coefficients of the sum of the education of both twins,

are larger than the return to education, and are significantly different from zero.

In the two next sub-sections, we conduct a series of sensitivity tests on the within-twin-

pair estimates. As with the conventional OLS estimates, the within-twin-pair estimates may

also be subject to biases that are caused by omitted variables and measurement errors.

4.4 Potential Biases of Within-Twin-Pair Estimates

Bound and Solon (1999) examine the implications of the endogenous determination of which

twin goes to school for longer, and conclude that twins-based estimation is vulnerable to the

same sort of bias that affects conventional cross-sectional estimation. The major concern of

the within-twin-pair estimate is thus whether it is less biased than the OLS estimate, and

is therefore a better estimate (Bound and Solon, 1999; Neumark, 1999). They argue that

although taking a within-twin-pair difference removes genetic variation, that is, it removes µi

from Equation (4), this difference may still reflect an ability bias to the extent that ability

consists of more than just genes. In other words, within-twin-pair estimation may not

completely eliminate the bias of conventional cross-sectional estimation, because the within-

twin-pair difference in ability may remain in 1i−2i in Equation (4), which may be correlated

with Z 1i − Z 2i. If endogenous variation in education comprises as large a proportion of the

remaining within-twin-pair variation as it does of the cross-sectional variation, then within-

twin-pair estimation is subject to as large an endogeneity bias as cross-sectional estimation.

Although within-twin-pair estimation cannot completely eliminate the bias of the OLS

estimator, it can tighten the upper bound on the return to education. Ashenfelter and

Rouse (1998), Bound and Solon (1999), and Neumark (1999) have debated the bias with



OLS and within-twin-pair estimation at length in recent papers. Note that the bias in the

OLS estimator depends on the fraction of variance in education that is accounted for by

variance in unobserved ability that may also affect earnings, that is, cov(Z i,µi+i)var(Z i)

. Similarly,

the ability bias of the fixed effects estimator depends on the fraction of within-twin-pair

variance in education that is accounted for by within-twin-pair variance in unobserved ability

that also affects earnings, that is, cov(∆Z i,∆µi+∆i)var(∆Z i)

. If we are confident that education and

the earnings error term are positively correlated both in the cross-sectional and within-

twin-pair regressions, and if the endogenous variation within a family is smaller than the

endogenous variation between families, then the fixed effects estimator is less biased than

the OLS estimator. Hence, even if there is an ability bias in the within-twin-pair regressions,

the fixed effects estimator can still be regarded as an upper bound on the return to education

(if education and ability are positively correlated). In that case, we can credit the within-

twin-pair estimates with having tightened the upper bound on the return to education.

To examine whether the within-twin-pair estimate is less biased than the OLS estimate,

we follow Ashenfelter and Rouse (1998) and conduct some correlation analyses. We use

the correlations of average family education over each twin pair with the average family

characteristics that may be correlated with ability (for example, marital status, spousal

education, membership of the Chinese Communist Party, working in a foreign firm, and job

tenure) to indicate the expected ability bias in a cross-sectional OLS regression. We then

use the correlations of the within-twin-pair differences in education with the within-twin-pair

differences in these characteristics to indicate the expected ability bias in a within-twin-pair

regression. If the correlations in the cross-sectional case are larger than those in the within-

twin-pair case, then the ability bias in the cross-sectional regressions is likely to be larger

than the bias in the within-twin-pair regressions.

The correlation tests that are reported in Table 3 suggest that the within-twin-pair



estimation of the return to education may indeed be less affected by omitted variables than

the OLS estimation. Note that the between-family correlations are all larger in magnitude

than the within-twin-pair correlations. For example, the correlation between average family

education and average spousal education is as large as 0.62 (column 1, row 2), which suggests

that twins in families with a high average level of education marry highly educated people.

This is consistent with the assumption that spousal education reflects an individual’s ability

and family background. The correlation of the within-twin-pair difference in education and

the within-twin-pair difference in spousal education is about a quarter of the between-family

correlation. This suggests that, to the extent that spousal education measures ability, the

within-twin-pair difference in education is less affected by ability bias than the average family

education. However, this within-twin-pair correlation is still statistically significant and large

in magnitude, which suggests that within-twin-pair differencing cannot completely eliminate

the ability bias that is embodied in education. Thus, the within-twin pair estimation may

only establish an upper bound for the estimated return to education. The correlations of

education with other variables provide similar evidence that the within-twin-pair estimation

is subject to a smaller omitted ability bias. Of course, these characteristics are only an

incomplete set of ability measures, but the evidence is suggestive.

4.5 Measurement Error

Another issue that we need to deal with is the measurement error problem. As is well

known, classical errors in the measurement of schooling lead to a downward bias in the

estimate of the effect of schooling on earnings, and the fixed effects estimator magnifies such

measurement error bias (Woodridge, 2002).

One way to solve the problem of measurement error bias is to use the instrumental

variable method. In this study, we follow the innovative approach of Ashenfelter and Krueger

(1994) to obtain good instrumental variables. More specifically, in our survey we asked each



twin to report both their own education and their co-twin’s level of education. In the presence

of measurement error in self-reported education, cross reported education is a potential

instrument, as the report of the other twin should be correlated with the true education

level of a twin but uncorrelated with any measurement error that might be contained in the

self-report.

Following Ashenfelter and Krueger (1994), the instrumental variable approach can be

applied as follows. Writing Z k j for twin k’s report of twin j’s schooling gives four different

ways to express the schooling difference between the two twins.

∆Z = Z 11− Z 2

2(8)

∆Z = Z 21 − Z 12 (9)

∆Z ∗ = Z 11 − Z 12 (10)

∆Z ∗∗ = Z 21 − Z 22 . (11)

Assuming classical measurement error, that is, that measurement error in each of these

reported education variables is uncorrelated with measurement error in the other variables,

we can fit

∆yi = ∆Z iβ + ∆i, (12)

using ∆Z as an instrument for ∆Z . This approach is valid even in the presence of common

family-specific measurement error, because the family effect is eliminated from both ∆Z

and ∆Z

. We call this instrumental variable model the IVFE-1.

The IVFE-1 estimates of Equation (12) that are reported in the first two columns of

Table 4 show that measurement error has biased downward the fixed effects estimates in

columns (5) and (6) of Table 2, as other studies in the literature. The IVFE-1 estimates of

the return rise by about 22 percent (from 0.027 by the fixed effects model to 0.033 by the

IVFE-1 model), which suggests that a considerable fraction of the variability in the reported

differences in the education levels of twins is due to measurement error. In other words,



the conventional fixed effects method is producing serious underestimates of the economic

returns to schooling.

However, the IVFE-1 estimates may also be biased if the measurement error terms in

∆Z

and ∆Z

are correlated. If there is an individual-specific component of the measurement

error in reporting education, then Z 11 and Z 12 will contain the same reporting error. As a

result, the error terms in ∆Z and ∆Z will be correlated, which makes ∆Z an invalid

instrumental variable for ∆Z .

Before discussing another instrumental variable, it is worth examining the correlations

between the educations variables and their correlation with earnings as reported in Table 5.

To facilitate this examination, we use the same correlations that are reported in Ashenfelter

and Krueger (1994) as benchmarks. Interestingly, the correlations in our sample are very

similar to those of Ashenfelter and Krueger. First, the correlations between the self-reported

and co-twin-reported education of the same twin, that is, cov(Z 11 , Z 21 ) and cov(Z 22 , Z

12), are

0.932 and 0.923 in our sample, compared to 0.920 and 0.877 in the sample of Ashenfelter

and Krueger. These high correlations suggest that the co-twin-reported level of education

is a good instrumental variable for self-reported level of education in our sample. Second,

our figures for the correlation between one twin’s self-reported education and his/her re-

port of the co-twin’s education, that is, cov(Z 11 , Z 12) and cov(Z 22 , Z

21), are 0.739 and 0.720,

whereas the same correlations in the paper of Ashenfelter and Krueger are 0.700 and 0.697.

These correlations suggest that our sample may suffer from a slightly more serious correlated

measurement error problem.

This correlated measurement error problem motivates us to implement a better in-

strumental variable that will be valid even in the presence of correlated measurement errors

(Ashenfelter and Krueger, 1994). To eliminate the individual-specific component of the

measurement error in the estimation, it is sufficient to use the schooling differences that are



defined in (10) and (11), that is the education difference as reported by each twin, with one

being used as an instrument for the other. More specifically, we can estimate the following

equation,

∆yi = ∆Z

∗

i β + ∆i, (13)

using ∆Z ∗∗ as an instrumental variable for ∆Z ∗.

The estimates of the IVFE model that allows for correlated measurement errors, which

we call the IVFE-2, are reported in the last two columns of Table 4. The new estimates of

the return to education are 3.6-3.8 percent, or about 15 percent greater than the IVFE-1

estimates. Since our sample has correlated measurement error, which is similar to the sample

of Ashenfelter and Krueger (1994), the IVFE-2 model should have the best estimates of the

return to education.

5 What Is Distinct about China?

It is interesting to compare our estimates to other estimates in the literature that draw on

data from different countries, mostly rich Western countries. Note first that our estimate of

the raw return to education, that is, the OLS estimate, is 8.4 percent (column 4 of Table 2),

which is very close to other estimates in the literature (first column of Table A1). However,

our within-twin-pair estimate is only 2.7 percent, which is smaller than most estimates in

the literature. Moreover, the ability bias in our sample, which stands at 5.7 percent, is larger

than the ability bias that has been found in all other studies. These results together suggest

that, in the Chinese case, most of the raw return to education is actually ability bias, which

is different from the findings from other countries.

To ascertain why the true return to education is so low and the ability bias so high in

China and to evaluate why China is so different from other countries, we need to understand

the distinct education system in China, because this education system may explain the high

ability bias and low return to education in these estimates.



5.1 The Chinese Education System

The Chinese education system is highly selective and exam oriented. It is composed of

two stages: the compulsory stage and the non-compulsory stage. The compulsory stage

comprises six years of primary school and three years of junior high school. Currently, most

urban children finish nine years of compulsory education. Junior high school graduates have

a choice of attending high school or vocational school,5 and are required to take an entrance

exam to gain a place at either type of institution. High school graduates are eligible for the

college entrance exam, but vocational school graduates are generally not. In our monozygotic

(MZ) twins sample, 74 percent had a high school or vocational school degree or above.

6

Because of the huge number of people waiting to be educated and the limited number of

places at colleges and universities, entrance to college is extremely competitive, and only 13

percent of the workers in our sample obtained a college degree. To select those who will go

on to college education, a nationwide college entrance exam system has been adopted, and

the exam days of June 7, 8, and 9 determine the future of many young people each year:7

Those who pass the examinations will become “white collar” workers, and those who fail

them will most likely become “blue collar” workers.

Because of the competitive nature of the education system, schools, and in particular

junior high schools and high schools, place great emphasis on exam-taking techniques.8

Although high school in China lasts for three years, the whole curriculum is normally finished

in one and a half years or an even a shorter time, with the rest of the time being spent on

preparation for the college entrance exams. Although the first half of high school teaches

5There are several types of vocational schools in China, which are called vocational schools, technicalhigh schools, or skilled workers’ schools. In this paper, we group them together under the term “vocationalschools.”

6The percentage for the whole sample is 72 percent.7The exam dates were formerly 7, 8, and 9 July, but were changed in 2003 to avoid the hot weather.8It is no secret that the Chinese have very good exam taking skills. For example, among graduate school

applicants in the United States, those from China normally have very high scores in GRE and other standardtests, and sometimes even have higher test scores in verbal English than native speakers. However, most

Chinese people have never spoken English before coming to the United States because oral English is neitherrequired by most US graduate schools nor emphasized in the English exams in China.



students new things, the teaching is also focused on exam-type problem-solving techniques.

High school students need to finish a lot of homework every day, and normally need to go to

school on weekends and vacations. All of this extra time is spent on training students to solve

exam questions. Schools and teachers are rewarded solely on the basis of the success rate

of their students in the entrance exams, and thus have no incentive to teach them anything

else. These exam-taking techniques very often have little to do with the knowledge and skills

that are needed for life and work, and it is thus unsurprising that such kind of schooling has

a low return in the workplace.

Two other features of the Chinese non-tertiary education system are also distinct.

First, the curricula ( jiao xue da gang in Chinese) for primary school, junior high school,

and high school are fixed by the Ministry of Education, and the most important part of

these curricula is to specify what should be covered by the high school and college entrance

exams. Schools and teachers then follow these curricula to prepare students for these exams.

Second, high school students have to decide to take either arts or science for the rest of

their education. Both arts and science students take Chinese, English, and political science,

but arts students take geography, history, and basic mathematics, whereas science students

take physics, chemistry, biology, and advanced mathematics. The college entrance exams

also have two sets of papers, one for arts and the other for science. Because of the fixed

curriculum, many students may not be able to study what they really are interested in,

and having to make an early decision on whether to specialize in arts or science prevents

students from obtaining the general education or training that are needed for life and work.

Moreover, young students have to decide what they want to do before they even know what

they are truly interested in, and as a result often choose badly.

Education that is not exam oriented only takes place in vocational schools and colleges.

First, as vocational schools or colleges are different from each other and are administered



by different ministries and provinces, they have the freedom to choose their own curricula.

Second, and most importantly, vocational schools and colleges are usually the final stage of

education, and thus exams are no longer important. Normally, vocational school graduates

are not allowed to take the college entrance exams (and thus have no chance of going to

college), and although college students may take the entrance exam to go on to graduate

school, only a small proportion of students choose to do so. As exams are not important

any more, students can spend more time on their true interests, and college students can

select courses in different departments, although changing one’s major (which is determined

during the college admission process) is still difficult.

The distinct features of the Chinese education system can help to explain why the

omitted ability bias (or selection effect) is high and the true return to education low in our

estimations. Because of intense competition, only the very talented can advance to higher

education, and thus education (or entrance exams) is a very good selection mechanism.

Because of the exam-oriented education system, non-tertiary education, and in particular

high school education, has little value-added in terms of general knowledge or workplace

skills, except as a means of selecting talented candidates into college. High school graduates

who are not able to get into college may thus have wasted three years on training in exam-

taking techniques.

5.2 What Levels of Education Pay?

The distinct education system not only helps to explain why the return is low and ability

bias high, but also suggests that the return to education may differ across education levels.

It seems that exam-oriented high school education is the least useful level of education, and

is valuable only as a selection mechanism for colleges. This means that the education that

high school graduates who do not make it to college obtain should be least rewarded by

employers. We investigate whether this is true by estimating the return to different levels of



education by allowing the return to education to vary across education levels.

In the literature on twins studies, years of schooling is generally used as the measure

of education (see, for example, Ashenfelter and Krueger, 1994; Ashenfelter and Rouse, 1998;

Bonjour et al., 2003). However, little work has been carried out to examine the returns

to different levels of education. As most of the literature draws on data from developed

countries, and a large proportion of workers in these countries have some years of college

education and have at least completed a high school education, it may not be necessary to

examine the return to high school education versus the return to college education. However,

knowing the returns to different levels of education is still very important for a developing

country such as China, where college education is very limited. Knowing the returns to

different levels of education could help the government to better allocate limited resources

for education.

In Table 6, we report the regressions using three education dummies as measures of

education. The high school dummy equals 1 if the last qualification that an individual

obtained was a high school qualification, and 0 otherwise. Similarly, the vocational school

dummy equals 1 if the last qualification that an individual obtained was a vocational school

qualification, and 0 otherwise. The college dummy equals 1 if an individual had a college

education or above, and 0 otherwise.

Our regression results show that it pays to attend vocational school or college, but

not to attend high school only. The high school dummy is positive and significant in the

two OLS estimations (columns 1 and 2 of Table 6), and high school graduates on average

earn 7.0-10.5 percent more than those without a high school degree. However, this premium

becomes almost zero and insignificant when we take the within-twin difference in columns

3-5. In contrast, the premiums that are associated with vocational school and college are

large. With the OLS estimates, the vocational school premium is 32.4-34.4 percent and the



college premium is as large as 61.1-62.3 percent. The positive premiums remain, although

they become smaller, even after we control for omitted ability bias by taking the within-twin

difference. For our best estimator, which is the IVFE-2 model, the estimated premiums

for vocational school and college are 22.0 and 40.0 percent, respectively. These estimates

mean that individuals with a vocational school degree earn 22.0 percent more and those with

a college degree 40.0 percent more than individuals without a vocational school or college

degree.

It is also interesting to calculate the return to each year of schooling for vocational

school and college. As vocational school education usually lasts for three years, the return

to each year of schooling is 7.3 percent. As it takes seven years (three years of high school

plus four years of college) to gain a college degree, the average return to each year of schooling

is only 5.7 percent. However, as high school has a zero return, the return to each year of

college education can be as high as 10 percent.

These estimates have important policy implications. First, exam-oriented high school

education has no return unless one also attends college. Given that vocational school is a

substitute for high school in China, it is thus rather risky for an individual to go to high

school, as the chance of getting into college is low. This partially explains why many children

from poor families choose to go to vocational school, even if they are eligible for high school.

With limited college places, exams may be the only possible mechanism to select students

into college, but there may still be ways to make high school education more useful. For

example, the decision about whether to specialize in arts or science could be postponed until

college, which would thus make high school education more well rounded. As one major

function of junior high school and high school education is to select students into the next

level of education, the government could consider shortening junior high school and high

school education, say, from three years each to two years each, and using the saved resources



to expand the provision of vocational schools or college.

6 Conclusion

In this paper, we empirically measure the return to education in urban China. By using

twins data to control for omitted ability and the family effect, we find that most of the

return to education that is estimated by the OLS model is actually due to the effects of

unobserved ability or family background. In other words, our findings seem to suggest that

the selection role of education is more important than the knowledge that is acquired at

school. The fixed effects estimate of the return to one year of education is only 2.7 percent,

and the part of the education premium that is due to unobservable ability is as large as

5.7 percent. Our sensitivity analyses suggest that the within-twin-pair estimates are biased

downward because of measurement error, and after the correction of this bias the estimated

return rises to 3.8 percent. We further show that the return to high school education is zero,

but that the returns to vocational school and college education are 22.0 percent and 40.0

percent, respectively.

The earlier findings of a low return to education in China, such as those of Byron

and Manaloto (1990) and Meng and Kidd (1997), have generated a great deal of interest

among economists. In the search for explanations of the low return, most economists have

turned to the remaining elements of the planned economy. We agree that the return was low

when the Chinese economy was under a planned regime and is likely to rise as the economy

becomes more market-oriented (see, for example, Heckman and Li (2004) and Zhang et al.

(forthcoming)), but argue that economic transition may not be the whole story. Because of

the distinct education system in China, the returns to the various levels of education are

different, and in particular we find that exam-oriented high school education has no return

in terms of earnings; rather, it merely serves as a selection mechanism for colleges. Few

previous studies have paid attention to the distinctive education system in China (or that



of other Asian countries) and its consequences for the return to education.

Knowing the true return to education has important policy implications. Our findings

show that the return to education is not universally low in China. The return to each

year of vocational school is 7.3 percent, and the return for each year of college education

is as high as 10.0 percent, both of which are comparable to estimates that draw on twins

data from other countries. Thus, our findings support the argument of Heckman (2003 and

2005) and Fleisher and Wang (2004) that investing in human capital is worthwhile in China.

However, our results also have some particular policy implications. Given that China has

limited resources to devote to education, it is important to identify educational priorities.

Our finding that the return to high school education is zero and that high school education

only serves as a college selection mechanism in urban China suggests that nothing would

be lost if high school education were shortened by one year. The resources saved could be

invested in levels of education that yield higher returns, such as vocational school or college,

or could be diverted to the provision of basic education in rural China, where many children

are not fortunate enough to obtain the nine years of compulsory education (Brown and Park,

2002).



27

Reference

Ashenfelter, Orley and Alan B. Krueger. 1994. Estimating the Returns to Schooling

Using a New Sample of Twins. American Economic Review 84(5): 1157-1173.

Ashenfelter, Orley and Cecilia Rouse. 1998. Income, Schooling and Ability: Evidencefrom a New Sample of Identical Twins. Quarterly Journal of Economics 113(1):

253-284.

Becker, Gary. 1993. Human Capital: A Theoretical and Empirical Analysis, with

Special Reference to Education. Chicago, IL: University of Chicago Press.

Behrman, Jere R. and Mark R. Rosenzweig. 1999. ‘Ability’ Biases in Schooling

Returns and Twins: a Test and New Estimates. Economics of Education Review 18(2):

159-167.

Behrman, Jere R., Mark R. Rosenzweig and P. Taubman. 1994. Endowments and the

Allocation of Schooling in the Family and in the Marriage market: The Twins

Experiment. Journal of Political Economy 102(6): 1134-1174.

Behrman, Jere R., Mark R. Rosenzweig and P. Taubman. 1996. College Choice and

Wages: Estimates Using Data on Female Twins. Review of Economics and Statistics

73(4): 672-685.

Behrman, Jere R. and P. Taubman. 1976. Intergenerational Transmission of Income

and Wealth. American Economic Review 66(2): 436-440.

Behrman, Jere R., P. Taubman, T. Wales, and Z. Hrubec. 1977. Inter- and

Intragenerational Determination of Socioeconomic Success with Special Reference to

Genetic Endowment and Family and Other Environment. Mimeo. University of

Pennsylvania.

Bian, Yanjie and John Logan. 1996. Market Transition and Persistence of Power: The

Changing Stratification System in Urban China. American Sociological Review 61(5):

739-758.

Bonjour, Borothe, Lynn F. Cherkas, Jonathan E. Haskel, Denise D. Hawkes and Tim

D. Spector. 2003. Returns to Education: Evidence from U.K. Twins. American

Economic Review 93(5): 1799-1812.

Bound, Jone and Gary Solon. 1999. Double Trouble: On the Value of Twins-Based

Estimation of the Return to Schooling. Economics of Education Review 18(2):

169-182.



28

Brown, Philip and Albert Park, “Education and Poverty in Rural China,” Economics of

Education Review 21 (2002): 523-541.

Byron, Raymond P. and Evelyn Q. Manaloto. 1990. Returns to Education in China.

Economic Development and Cultural Change 38: 783-796.

Fleisher, Belton M. and Xiaojun Wang. 2004. Skill Differentials, Return to Schooling,

and Market Segmentation in a Transition Economy: The Case of Mainland China.

Journal of Development Economics 73(1): 315-328.

Gerber, Theodore and Micheal Hout. 1998. More Shock than Therapy; Employment

and Income in Russia, 1991-1995. American Journal of Sociology 104(1): 1-50.

Gorseline, D. W. 1932. The Effect of Schooling upon Income. Bloomington: Indiana

Univ. Press.

Heckman, James. 2003. China’s Investment in Human Capital. Economic

Development and Cultural Change 51(4): 795-804.

Heckman, James. 2005. China’s Human Capital Investment. China Economic Review

16: 50-70.

Heckman, James and Xuesong Li. 2004. Selection Bias, Comparative Advantage and

Heterogeneous Returns to Education: Evidence from China in 2000. Pacific

Economic Review 9(3): 155-171.

Herrnstein, Richard J. and Charles Murray. 1994. The Bell Curve. New York: FreePress.

Hertz, Thomas. 2003. Upward Bias in the Estimated Returns to Education: Evidence

from South Africa. American Economic Review 93(4):1354-68.

Isacsson, Gunnar. 1999. Estimates of the Return to Schooling in Sweden from a large

sample of twins. Labour Economics 6(4): 471-489.

Lam, David and Robert F. Schoeni. 1993. Effects of Family Background on Earnings

and Returns to Schooling: Evidence from Brazil. Journal of Political Economy 101(4):

710-740.

Meng, Xin and Micheal P. Kidd. 1997. Labor Market Reform and the Changing

Structure of Wage Determination in China’s State Sector during the 1980s. Journal of

Comparative Economics 25(3): 403-421.

Miller, P., C. Mulvey and N. Martin. 1995. What Do Twins Studies Reveal About the



29

Economic Returns to Education? A Comparison of Australian and U.S. Finding.

American Economic Review 85(3): 586-599.

Mincer, Jacob. 1974. Schooling, Experience and Earnings. New York: Columbia

University Press.

Myers, R.H.. 1995. Chinese Debate on Economic Reform: Can China Create A

Socialist Market Economy? Asian-Pacific Economic Literature 9(2): 55-68.

Neumark, David. 1999. Biases in Twin Estimates of the Return to Schooling.

Economics of Education Review 18(2):143-148.

Parish, William L. and Ethan Michelson. 1996. Politics and Markets: Dual

Transformations. American Journal of Sociology 101(4): 1042-1059.

Psacharopoulos, George. 1992. Returns to Education: A Further International Update

and Implications. International Library of Critical Writings in Economics 17:

102-123.

Rona-Tas, Akos. 1994. The First Shall Be Last? Entrepreneurship and Communist

Cadre in the Transition from Socialism. American Journal of Sociology 100(1): 40-69.

Rouse, Cecilia. 1999. Further Estimates of the Economic Return to Schooling from a

New Sample of Twins. Economics of Education Review 18(2): 149-157.

Szelenyi, Ivan and Eric Kostello. 1996. The Market Transition Debate: Toward a

Synthesis. American Journal of Sociology 101(4): 1082-1096.

Taubman, Paul. 1976a. The Determinates of Earnings: Genetics, Family and Other

Environments, a Study of White Male Twins. American Economic Review 66(5):

858-870.

Taubman, Paul. 1976b. Earnings, Education, Genetics, and Environment. Journal of

Human Resources 11(4): 447-461.

Walder, Andrew. 1996. Markets and Inequality in Transitional Economics: Toward

Testable Theories. American Journal of Sociology 101(4): 1060-1073.

Woodridge, Jeffrey. 2002. Econometric analysis of cross section and panel data.

Cambridge, Massachusetts: MIT Press.

Xie, Yu and Emily Hannum. 1996. Regional Variation in Earnings Inequality in

Reform-Era Urban China. American Journal of Sociology 102(4): 950-992.



30

Zhang, Junsen, Yaohui Zhao, Albert Park and Xiaoqing Song. Economic Returns to

Schooling in Urban China, 1988 to 2001. Forthcoming in Journal of Comparative

Economics.

Zhou, Xueguang. 2000. Economic Transformation and Income inequality in Urban

China. American Journal of Sociology 105(4): 1135-1174.



31

Table 1: Descriptive Statistics of the Twins and Non-Twins Samples

Variable MZ twins

(1)

All twins

(2)

Non-twins

(3)

NBS sample

(4)

Education (years of schooling) 12.22 12.16 11.73 11.62

(2.89) (2.91) (3.07) (2.83)

High school dummy 0.27 0.25 0.30 --

(0.44) (0.43) (0.46) --

Technical school dummy 0.34 0.35 0.32 --

(0.47) (0.48) (0.47) --

College dummy 0.13 0.12 0.10 --

(0.33) (0.33) (0.30) --

Age 34.78 33.77 43.27 40.80

(9.64) (9.22) (8.42) (11.98)

Gender (male) 0.60 0.59 0.48 0.55

(0.49) (0.49) (0.50) (0.50)

Married 0.66 0.64 0.94 --

(0.47) (0.48) (0.24) --

Tenure (the number of years in full-time work 15.03 14.03 21.70 18.45

since the age of 16) (9.93) (9.50) (9.05) (12.94)

Earnings (monthly wages, bonuses, and subsidies 887.85 872.52 845.84 1062.92

in RMB) (517.91) (546.00) (549.08) (840.09)

Spousal education 11.64 11.69 11.49 --

(3.11) (3.08) (3.49) --

Sample size 976 1620 1277 23288

Note: The mean and standard deviation (in parentheses) are reported in the table. For the MZ twins sample, we

restrict the sample to those twin pairs (488 pairs) for which we have complete information on earnings, age,

gender, years of education, job tenure, and marital status for both twins in the pair. The NBS sample is based

on a large-scale survey by the National Bureau of Statistics in six provinces.



32

Table 2: OLS and Fixed Effects Estimates of the Return to Education for Twins and Non-twins from Urban China

(Dependent variable: log earnings)

Sample Twins and non-twins Twins Twins Twins

Model OLS OLS FE GLS

(1) (2) (3) (4) (5) (6) (7) (8)

Own education 0.067*** 0.067*** 0.082*** 0.084*** 0.025* 0.027* 0.025** 0.027**

(16.71) (16.91) (13.85) (14.14) (1.68) (1.87) (2.06) (2.22)

Sum of 0.033*** 0.033***

education (4.70) (4.66)

Age 0.023** 0.011 0.041*** 0.036* 0.039*** 0.033**

(2.49) (0.89) (2.60) (1.88) (2.71) (2.00)

Age squared -0.020* -0.023* -0.045** -0.052** -0.040** -0.047**

(1.68) (1.67) (1.99) (2.13) (2.04) (2.26)

Gender (male) 0.217*** 0.210*** 0.205*** 0.202*** 0.206*** 0.204***

(9.05) (8.72) (5.32) (5.25) (5.36) (5.30)

Married -0.033 -0.027 -0.043 -0.025

(0.75) (0.53) (0.83) (0.58)

Tenure 0.016*** 0.011* 0.015 0.012**

(4.77) (1.86) (1.52) (2.09)

Twin pairs 488 488 488 488

Observations 2255 2253 976 976 976 976 976 976

R-square 0.17 0.18 0.22 0.23 0.01 0.02

Note: All of the OLS regressions include city dummies. Robust t statistics are in parentheses. * significant at 10%; ** significant

at 5%; *** significant at 1%.



33

Table 3: Between-Families and Within-Twin-Pair Correlations of Education and Other Variables (488

twin pairs)

Between-family correlations Within-twin-pair correlations

Education ∆Education

Married -0.1445*** ∆Married -0.0173(<0.01) (0.70)

Spousal education 0.6172*** ∆Spousal education 0.1518**

(<0.01) (0.02)

Party member 0.2571*** ∆Party member 0.1166**

(<0.01) (0.02)

Working in foreign firm dummy 0.0904* ∆Working in foreign firm dummy 0.0214

(0.06) (0.66)

Tenure -0.2614*** ∆Tenure -0.1253***

(<0.01) (0.01)

Note: The significance levels are in parentheses. * significant at 10%; ** significant at 5%; *** significant at

1%. The between-family correlations are the correlations between average family education (average of the

twins) and average family characteristics, and the within-twin-pair correlations are the correlations between

the within-twin-pair differences in education and the within-twin-pair differences in other characteristics.



34

Table 4: Instrumental Variable Fixed Effects Estimates of the Return to Education of Chinese Twins (Dependent

variable: log earnings)

IVFE-1

(∆Z’’ as IV)

IVFE-2

(∆Z** as IV)

(1) (2) (3) (4)

Education (∆Z’) 0.032* 0.033*

(1.65) (1.77)

Education (∆Z*) 0.036** 0.038**

(1.99) (2.14)

Married -0.043 -0.048

(0.81) (0.91)

Tenure 0.016 0.016

(1.60) (1.59)

Twin pair 488 488 488 488

Observations 976 976 976 976

Note: ∆Z’ is the difference between the self-reported education of twin 1 and the self-reported education of twin 2.

∆Z’’ is the difference between the education of twin 1 as reported by twin 2 and the education of twin 2 as

reported by twin 1. ∆Z* (∆Z**) is the difference between twin 1’s (twin 2’s) report of his/her own education

and his/her report of the other twin’s education. The robust t-statistics are in parentheses. * significant at 10%;

** significant at 5%; *** significant at 1%.



35

Table 5: Correlations Between Earnings and the Education Variables

Variable1Y 2Y

1

1 Z 2

1 Z 2

2 Z 1

2 Z 1

F E 2

F E 1

M E 2

M E

1Y 1.000

2Y 0.506 1.000

1

1 Z 0.388 0.375 1.000

2

1 Z 0.397 0.373 0.932 1.000

2

2 Z 0.366 0.417 0.758 0.720 1.000

1

2 Z 0.359 0.401 0.739 0.699 0.923 1.000

Father’s education

(1

F E )0.140 0.159 0.322 0.327 0.394 0.397 1.000

Father’s education

(2

F E )

0.128 0.158 0.315 0.323 0.394 0.396 0.979 1.000

Mother’s education

(1

M E )0.152 0.120 0.278 0.296 0.341 0.345 0.554 0.539 1.000

Mother’s education

(2

M

E )0.141 0.113 0.278 0.275 0.348 0.346 0.540 0.525 0.986 1.000

Notes: 1Y and 2Y represent twin 1’s and twin 2’s log monthly wage rate, respectively.k

j Z represents twin k’s

report of twin j’s education, where k = 1, 2 and j = 1, 2.k

F E (k=1, 2) represents the father’s education as reported

by twin k, andk

M E (k = 1, 2) represents the mother’s education as reported by twin k.



36

Table 6: Various Estimates of the Return to High School and College Education for Twins and Non-twins from

Urban China

Dependent variable: log earnings

Sample All Twins Twins Twins Twins

Model OLS OLS FE IVFE-1 IVFE-2(1) (2) (3) (4) (5)

High school 0.070** 0.105* -0.003 0.030 0.031

(2.19) (1.94) (0.04) (0.31) (0.29)

Technical school 0.324*** 0.344*** 0.168** 0.218** 0.220*

(10.40) (6.72) (2.09) (2.22) (1.89)

College 0.611*** 0.623*** 0.278** 0.392*** 0.400***

(16.32) (10.88) (2.45) (2.96) (2.66)

Age 0.022* 0.055***

(1.78) (2.76)

Age squared -0.038*** -0.075***

(2.70) (3.03)

Gender (male) 0.204*** 0.189***

(8.43) (4.73)

Married -0.049 -0.052 -0.039 -0.035 -0.040

(1.14) (0.99) (0.75) (0.66) (0.76)

Tenure 0.016*** 0.008 0.014 0.015 0.014

(4.68) (1.17) (1.42) (1.53) (1.48)

Twin pairs 488 488 488

Observations 2253 976 976 976 976

R-square 0.18 0.18 0.03

Note: All of the OLS regressions include city dummies. For model IVFE-1, we use∆Z’ (the difference

between the high school and college dummies) as independent variables, which are instrumented by∆Z’’.

For model IVFE-2, we use ∆Z* as independent variables, which are instrumented by ∆Z**. The robust

t-statistics are in parentheses. * significant at 10%; ** significant at 5%; *** significant at 1%.



Table A1: Estimated Return to Years of Education Using Twins Sample from Different Countries

Study Sample and Country OLS

(A)

FE

(B)

Omitted

variable

bias(C=A-B)

IVFE

(D)

Taubman (1976a) NAS-NRC Twin Registry sample

of white male army veterans, USA

0.079 0.027 0.052 --

Ashenfelter and Krueger

(1994)

Twinsburg sample, USA 0.084 0.092 -0.008 0.129

Behrman et al. (1994) NAS-NRC Twin Registry,

Minnesota Twin Registry, USA

-- 0.035 -- 0.050

Miller et al. (1995) Australia Twin Registry 0.064 0.025 0.039 0.048

Behrman et al. (1996) Male twins born in Minnesota,

USA

-- 0.075 -- --

Ashenfelter and Rouse

(1998)

Twinsburg sample, USA 0.110 0.070 0.040 0.088

Behrman and Rosenzweig(1999)

Minnesota Twin Registry, USA -- -- -- 0.104

Rouse (1999) Twinsburg sample, USA 0.105 0.075 0.030 0.110

Isacsson (1999) Sweden Twin Registry 0.049 0.023 0.026 0.024

Bonjour et al. (2003) Twins Research Unit, St., Thomas’

Hospital (female only), London,

UK

0.077 0.039 0.038 0.077

This study Chinese Twins Survey, China 0.084 0.027 0.057 0.038

Does Education Pay in Urban China- Estimating Returns to Education Using Twins

Documents